Researched
Machine Translation
From IBM and Georgetown's 1954 demo of 60 Russian sentences to Google's neural system of 2016, computers learn to translate between human languages.
Open in the interactive tree →Warren Weaver's 1949 memorandum, which drew on wartime codebreaking, started the field. In January 1954 IBM and Georgetown translated about 60 Russian sentences with 250 words and six rules and predicted a solution within three to five years; the 1966 ALPAC report found no useful progress and funding was cut sharply. Statistical methods that learn from parallel texts took over from the late 1980s, Google Translate launched in 2006 on UN and EU documents, and neural translation with attention (2014) replaced them in 2016, cutting errors by about 60% on simple sentences in Google's tests.
Prerequisites
- Programmable Computer (Zuse Z3)1941Translation needs a machine that can store a dictionary and rules.
- Information Theory1948Weaver's 1949 memorandum treated translation like decoding a message.
- Formal Grammars (Chomsky)1956Rule-based systems write grammar as formal rules.
Unlocks
- Attention Mechanism2014Attention was invented to let a neural translation system look back at the source words.
- AI for Low-Resource Languages2022-2026NLLB and the added Google Translate languages are machine-translation systems.